Abstract
While scientific applications can gather consistent data from the natural world, psychological, sociological, and even economic applications rely on data provided by people. Since the majority of machine learning is aimed at improving the lives of people, human input is essential for useful results. In this paper, we explore datasets where input and target attributes are provided by people taking surveys. Every survey dataset, generated from human input, is reliable and self-consistent according to Cronbach's alpha. One expects a reliable questionnaire to provide effective data for learning. It is this expectation that our analysis finds false, when applied to supervised learning. Both statistical analysis and application of several supervised learning architectures, with a focus on neural networks, are utilized to provide insight into data gathered through human input.